Neural networks for increasing the resolution of computational fluid dynamics simulation results with coarse meshes

• Grant number: 24/10776-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): November 01, 2024
• Effective date (End): October 31, 2025
• Field of knowledge: Engineering - Aerospace Engineering - Aerodynamics
• Principal Investigator: Gabriel Pereira Gouveia da Silva
• Grantee: Henrique Babo Terra
• Host Institution: Faculdade de Engenharia. Universidade Estadual Paulista (UNESP). Campus Experimental São João da Boa Vista. São João da Boa Vista , SP, Brazil

Abstract

Fluid flows play a vital role in several areas, including aerospace, automobile, marine, chemical, medical, and the study of pollutant dispersion. The movement of fluids is described by the Navier-Stokes equations, which do not have analytical solutions for most cases, with most studies relying on experiments or numerical simulations known as computational fluid dynamics (CFD). The numerical resolution of these equations is intensive regarding computational resources, with costs increasing according to the complexity of the geometries and the refinement of the domain discretization meshes. Although less refined meshes provide some results, they are not sufficiently accurate, making mesh convergence analysis essential to determine the smallest possible mesh that still produces accurate results. Recently, machine learning models have been used to improve CFD results, whether through adaptive meshes, turbulence closure models, or reduced-order surrogate modeling, among other possibilities. One promising application is the use of super-resolution or upscaling algorithms, which infer high-resolution data from low-resolution data, using statistical patterns extracted from high-resolution images during training. This technique, originating from image processing, can be adapted to flow field data from CFD simulations, allowing detailed, high-resolution data to be obtained from less refined and faster simulations. In this context, this project aims to create a database with results from two-dimensional computational fluid dynamics simulations for different geometries and mesh refinement levels. Using this database, machine learning models capable of predicting high-resolution results from simulations performed with coarse meshes will be trained and optimized. The simulations will be performed using Ansys Fluent software, focusing on two-dimensional symmetric geometries to maximize mesh refinement within academic license limitations. The flow field data obtained from the more refined meshes will serve as a reference for training the machine learning models. Several machine learning techniques will be tested and the best approach will be optimized in terms of its hyperparameters to provide high-resolution results from simulations made with coarse meshes. These coarse-mesh simulations and subsequent upscaling will then be compared regarding the duration and accuracies with conventional simulations using refined meshes in Ansys Fluent.